Claude Code permissions for teams: Plan, Manual and Auto
A team needs to decide what a coding agent may do without asking, and who answers when it does ask. In Friday every session runs in one of three modes, Plan, Manual or Auto, and approvals are visible to everyone in the session.
What is Friday Work?
Friday Work is a multiplayer AI workspace for coding teams. Your people and your coding agents work in one live session: the same task, the same record, the same tools. It is a desktop app for Windows and Mac.
- Teammates join a session by link, watch every step, suggest changes and take control when it is their turn.
- Claude Code, Codex and Cursor run on your own machine and your own accounts, side by side in one session, with Friday’s own agent when you want it.
- Continue a Claude Code or Codex session you already have, instead of starting over.
- Plan, Manual or Auto: you decide what an agent may do, and approve its commands as a team.
- Connect GitHub, Slack, Linear and 50+ more tools over MCP once, and choose which agents may use them.
- Free to start. Pro is $20 a month ($18 a month billed yearly).
What are the three modes?
| Mode | What the agent may do | Use it when |
|---|---|---|
| Plan | Read the code and propose a plan. It changes nothing until the driver approves the plan | You want to agree the approach before any file changes |
| Manual | Edit files. Commands wait for approval, and the approval shows exactly what it wants to run | Everyday work where you want a person in the loop |
| Auto | Edit files and run commands without stopping to ask | A task you trust, in a folder you can reset |
The mode applies to the agent you are running, whether that is Claude Code, Codex, Cursor or Friday’s own agent.
Who approves a command?
Each agent has one driver at a time. An approval goes to the driver of the agent that asked, and the session record shows who approved what and when. Teammates watch the same approval card, so a second pair of eyes sees what is about to run before it does.
A teammate can ask for control, or suggest a redirect, but they cannot make an agent run something the driver has not approved.
How should a team set it up?
- Start new work in Plan, and read the plan together.
- Move to Manual to edit. Approve commands one at a time, and read what each one does.
- Use Auto only for a bounded task in a folder you can reset, and check the receipt afterwards.
- Keep one person as the driver of each agent.
After a run, the receipt lists the commands it really ran and whether they passed, and flags a test that was weakened or a pass that never ran.
Common questions
Do the modes work with Claude Code, Codex and Cursor, not only Friday’s own agent?
Yes. Plan, Manual and Auto all work with Claude Code, Codex and Cursor, so the team gets the same approvals whichever tool is running.
Can a viewer approve a command?
No. Approvals go to the driver. Viewers can watch.
Is there a record of approvals?
Yes. The session log records who asked, what the agent did, and who approved it.